The AI mistakes that hurt an online shop most are product copy that invents specifications, descriptions so alike that shoppers can't choose, images that misrepresent the item, chatbots that promise what you don't offer, and reviews that aren't genuine. Each one raises returns or abandoned baskets, and most can be fixed with a checking step, not by dropping AI.
The damage rarely shows up as a single disaster. It shows up as a slow drift: returns tagged "not as described" creeping from 3% to 7%, more pre-sale questions about things the page should answer, a conversion rate that slips a few tenths of a percent after a catalogue rewrite. That drift is easy to blame on the market. Below are the eight mistakes worth ruling out first, each with an example of how it looks in a real shop.
The running example is illustrative: a plumbing firm that sells spare parts online alongside its call-out work. Tap cartridges, shower hoses, radiator valves, washers and seals, about 600 products. Parts shops are unforgiving of AI errors, because a part either fits or it doesn't.
1. Product copy that invents specifications
What goes wrong. Ask a chatbot to "write a product description for a ceramic disc tap cartridge" with only the product name, and it fills the gaps with plausible details. Plausible is not accurate.
How it looked. The firm fed its product names into a chatbot to fill 200 empty descriptions. One cartridge came back as "compatible with most standard mixer taps, quarter-turn, 20-spline". The real part was 24-spline and fitted two specific tap ranges. Within a month, returns on that line reached one in five, each one a customer who had drained down their system, taken the tap apart and found the part didn't fit. That is not a return; that is a customer who will never buy from you again.
The fix. Give the AI facts and forbid it to add any. A prompt that works:
Write a product description using ONLY these facts. If a detail
is missing, write [MISSING] instead of guessing. Never say
"universal", "fits most" or "compatible with all".
Facts: [product name, dimensions, spline count, turn type,
material, the exact tap ranges or model numbers it fits, what's
in the pack]
Structure: one line saying what it is and what it fits, then
bullet points for specs, then "Check before ordering:" with the
measurement the buyer should take.Fed the facts for the 24-spline cartridge, an illustrative reply looks like this: "Ceramic disc tap cartridge, 24-spline, quarter-turn, for [MISSING: tap ranges]. A reliable choice for most bathroom and kitchen taps. Specs: brass body, 35mm long, pack of one. Check before ordering: count the splines on your old cartridge." The [MISSING] marker did its job, because the fit list wasn't in the facts, and someone now has to look it up rather than the AI inventing it. But the second sentence slipped through the "fits most" ban in different words. Delete it, add a line to the prompt ("Don't describe who or what the product suits beyond the fit list"), and run the batch again. Catching one phrase like this in the first ten outputs is the reason to check a sample before running all 200.
Watch out for supplier data sheets that are themselves wrong or out of date. AI copies errors faithfully. Spot-check the first 20 descriptions against the physical part.
2. Every description sounds the same
What goes wrong. Run a whole catalogue through one prompt and every page opens the same way. Shoppers comparing three similar products can't see why they differ, so they leave to check elsewhere, or buy the cheapest and return it.
Before (three shower hoses, AI-written): "Upgrade your bathroom with this high-quality shower hose, designed for durability and style." The same sentence, give or take a word, on all three.
After (edited): "1.25m stainless steel hose, standard 1/2-inch fittings both ends. Choose this one for a fixed-height rail; choose the 1.75m if your rail slides above head height." The difference between products is now the first thing a shopper reads.
The fix. Make the prompt lead with what distinguishes each product from its nearest neighbours, and add a short comparison table on category pages. For the craft of copy that sells, see writing product descriptions with AI that actually sell; on Shopify specifically, Shopify product descriptions with AI that convert covers the platform's own tools.
3. Images that misrepresent the product
What goes wrong. AI image tools can place a product in a styled bathroom or generate a "lifestyle" shot from a plain photo. In doing so they often alter the product itself: the finish, the proportions, the number of holes.
How it looked. A generated image of a basin mixer showed a polished chrome finish; the product was brushed nickel. Another showed a radiator valve noticeably larger relative to the pipe than it really is. "Not as pictured" returns followed, and two reviews mentioned it by name.
The fix. Set a house rule: AI may change backgrounds and lighting, never the product. Keep at least one unedited photo of the real item on every page, first in the gallery for technical parts. Before publishing any generated image, compare it with the product side by side; checking AI-generated images for errors before posting has a short checklist.
Variants cause a quieter version of the same problem. A basin mixer sold in chrome, matt black and brushed brass gets one attractive AI bathroom shot, generated from the chrome photo and set as the main image for all three. A shopper picks matt black from the drop-down, the picture stays chrome, and she either abandons the basket or orders unsure. Most shop platforms let you attach an image to each variant so the picture changes with the choice; use a real photo of each finish there, and keep the styled shot further down the gallery.
Watch out for tools that "enhance" photos automatically when you upload them. Check the settings on your shop platform's image features.
4. A chatbot that makes promises you don't keep
What goes wrong. A shop chatbot without firm rules answers what it is asked, and customers ask about delivery dates, stock, compatibility and refunds. It guesses when it doesn't know.
How it looked (illustrative transcript):
Customer: If I order this valve now will it come tomorrow? Need
it before the weekend.
Bot: Yes! Orders placed today are delivered next day, so it
will arrive tomorrow.The shop's cut-off was 1pm and it was 3:40pm. The valve arrived two days later, the customer had already bought one locally, and the return and a one-star review followed. The bot wasn't malicious; it had read "next-day delivery available" on the shipping page and ignored the cut-off in the small print.
The fix. Give the bot live data (order status, stock, the current time against cut-off) or forbid it from promising anything it can't check: "Never promise a delivery date. Quote the dispatch rule and the courier's stated timescale." Whether an AI chatbot can handle order tracking and returns goes through the data connections and rules in detail. Remember that whatever your bot says, customers treat as your word, and so may a court.
5. Reviews that aren't what they seem
What goes wrong. Three versions of this, from worst to subtlest:
- AI-written fake reviews to fill empty product pages. Consumer-protection regulators in several markets now explicitly ban fake reviews, including AI-generated ones, and shoppers are increasingly good at spotting them. Never do this.
- AI "polishing" real reviews before publishing them. Once the words aren't the customer's, it isn't their review. Where the legal line sits for AI and testimonials covers this in more depth.
- AI review summaries that overstate. Some platforms and apps now generate a summary above the reviews.
How the last one looked. A summary on a push-fit connector read: "Customers love how easy it is to fit and say it's leak-free." Of 40 reviews, 29 were positive, but 8 mentioned slow leaks on older copper pipe. The summary dropped the one thing a careful buyer needed to know. When a shopper found those reviews further down, the summary made the whole page look untrustworthy.
The fix. Read the generated summary against the reviews at least monthly for your top 20 products, and switch summaries off where they can't be trusted to include the main complaint. Reply to the critical reviews too; a shop that answers "use the copper insert provided for older pipe" turns a warning into a helpful note.
6. Hiding the bot, or hiding the human
What goes wrong. A chatbot that presents itself with a human first name and a staff photo, and a support flow with no route to a person both erode trust the moment the customer realises. The first is also a compliance issue: if you sell to customers in the EU, the AI Act's transparency duty, in force since 2 August 2026, requires people to be told they are interacting with an AI.
How it looked. A customer with a mis-delivered order typed "can I speak to a person" four times and got four variations of "I'm here to help! What's your order number?" The chat transcript ended up quoted in a review.
The fix. An opening line that says what it is ("I'm the shop's automated assistant; I can check orders and answer product questions, and a person can take over at any time"), and a rule that "person", "human" or "agent" triggers a handover straight away. Suggested wording is in AI chatbot disclosure: what to tell customers at the start of a chat.
7. Personalisation that misfires
What goes wrong. AI-driven emails and "recommended for you" blocks work from purchase history, and purchase history doesn't know about returns, gifts, trade buyers or context.
How it looked. A customer returned a thermostatic shower valve because it didn't fit her pipework. Two weeks later she got an email: "Loving your new shower valve? Complete the look with these matching accessories." In the same month, a trade customer who bought 30 radiator valves for a job received "It's time to replace your radiator valves!" reminders a year later, for a house he'd never lived in.
The fix. Exclusion rules before any automated email goes out:
- Exclude anyone who returned or complained about the product the email refers to.
- Exclude trade or bulk buyers from consumer reorder flows.
- Only recommend accessories marked as compatible in your product data, not ones that are merely "frequently bought together".
- Cap automated emails per customer per month.
8. Machine translation nobody checked
What goes wrong. If you sell in more than one language, AI translation is fast and mostly good, and the "mostly" is the problem. Technical vocabulary is where it slips.
How it looked. A translated page rendered "olive" (the small brass sleeve in a compression fitting) as the fruit, and "isolating valve" as a word meaning an insulating valve. A plumber reading it would lose confidence in everything else on the site.
The fix. Give the translator a glossary of your trade terms with approved translations, and have a fluent speaker check the category pages, the checkout, the returns policy and your 20 best-selling products. The long tail can wait; those pages carry most of the trust.
A glossary doesn't need to be long to work. The first rows for this shop might read:
| Term | What it means here | Instruction to the translator |
|---|---|---|
| Olive | Brass ring that seals a compression fitting | Use the trade term for a compression ring; never the fruit |
| Isolating valve | Small valve that shuts off water to one tap or appliance | Not "insulating"; use the term plumbers use |
| Push-fit | Joint that connects by pushing the pipe in, no tools | Keep as the trade name if that's what buyers search for |
| Spline | The ridges on a cartridge spindle that the handle grips | Keep the number next to it exactly as written |
Paste the glossary at the top of every translation prompt, and ask the tool to list any term it wasn't sure how to translate. That list is where your fluent checker should start.
An illustrative trust audit, filled in
For the parts shop above, a two-hour audit a quarter after its AI rewrite might produce a sheet like this. Numbers are illustrative.
| Check | What we found | Action |
|---|---|---|
| Returns tagged "not as described", last 90 days | 7.1%, up from 3.4% before the rewrite; 60% on cartridges and valves | Re-check every cartridge and valve description against the part (mistake 1) |
| Pre-sale questions by email and chat | Top question: "will this fit my tap?" (41 times) | Add a "Check before ordering" line and fit list to each page |
| Images | 9 generated images alter the product finish or size | Replace with real photos; AI backgrounds only |
| Chatbot transcripts, sample of 30 | 4 delivery promises it couldn't check; 1 ignored request for a person | Add cut-off rule; add handover trigger words |
| Review summaries, top 20 products | 3 omit a recurring complaint | Switch summaries off on those; reply to the complaints |
| Automated emails | Reorder flow includes returners and trade buyers | Add exclusion rules |
Notice where the audit starts: with returns reasons and customer questions, not with the AI. Those two data sources tell you which pages are failing shoppers, and the fixes follow. In this example, the likely order of payback is descriptions first (the biggest returns cost), then the chatbot rules (cheap and quick), then images.
How to tell whether the fixes worked
Pick the three measures that match the mistakes you fixed and compare the 60 days after with the 60 days before, against the same period last year if your trade is seasonal:
- "Not as described" returns as a share of orders, by category.
- Pre-sale questions per 100 orders. Good product pages reduce them.
- Conversion rate on the pages you changed, compared with pages you left alone. If both move together, something else (a promotion, the season, a change in ads) is responsible.
For the parts shop, an illustrative comparison two months after the fixes:
| Measure | 60 days before | 60 days after | Reading |
|---|---|---|---|
| "Not as described" returns, cartridges and valves | 12.2% | 4.6% | The fit lists are working |
| "Not as described" returns, rest of shop | 4.4% | 4.0% | Little change, as expected |
| Pre-sale questions per 100 orders | 18 | 11 | "Will this fit?" questions fell the most |
| Conversion, changed pages | 2.1% | 2.4% | Up |
| Conversion, unchanged pages | 1.9% | 2.2% | Also up: the season, not the fix |
The last two rows are the lesson. Conversion rose everywhere by a similar amount, so the fixes can't claim it. The returns and questions rows moved only where the fixes were made, which is the evidence that they worked. On about 900 cartridge and valve orders over the period, returns falling from 12.2% to 4.6% means roughly 70 fewer customers draining down a system for a part that didn't fit.
None of these mistakes is an argument against using AI in an online shop. Every one comes from letting AI publish without a person who knows the products checking it. Put that person back into the loop at the right points, and AI saves the time without costing the trust.
Further reads
- Your First 30 Days of AI Customer Support for an Online Shop — A calmer first month if you are adding AI support now.
- Can a Small Online Shop Use AI to Compete With Bigger Brands? — Where AI genuinely helps a small shop against bigger brands.
- Shopify Magic vs ChatGPT for Product Descriptions — Which writing tool suits product copy on Shopify.
- Can You Legally Use AI-Generated Images in Your Marketing? — The legal side of AI images in product and marketing use.
- How to Stop an AI Chatbot Giving Customers Wrong Answers — Guardrails for a shop chatbot that keeps guessing.
- How to Build a Brand Voice Guide That AI Can Follow — Stop AI copy making your shop sound generic.
- How to Measure Whether Your AI Chatbot Is Actually Working — Numbers that show whether the bot helps or harms sales.
- AI Mistakes That Damage Customer Trust, and How to Avoid Them — Nine AI mistakes customers notice, why each one stings, how to prevent it, and a 20-minute monthly check that catches problems before customers do.
- How Much Does It Cost to Add AI to a Shopify Store? — Adding AI to Shopify can cost $0 extra, or $35 to $500+ a month with apps. Itemised prices, three store budgets and the costs listings don't show.
- Taking a Local Shop Online With AI: Product Listings in a Weekend — A Friday-to-Sunday plan for getting a shop's first 30-80 products online, with the spreadsheet, prompt and checks that keep AI listings accurate.
- How to Set Up AI Abandoned Cart Emails for a Small Online Store — Turn on the free built-in checkout reminder, add two follow-ups, and let AI write the copy while plain rules decide who is emailed and when.
- How to Write eBay Listings With AI That Sell Faster — Sold-price research, a title formula, item specifics and flaw-honest descriptions for eBay listings drafted with AI, plus the policy traps to avoid.
- AI Model Shots for Boutiques: Try-On Images Without a Photoshoot — How AI on-model images work, what tools like FASHN and Photoroom cost, which garments come out wrong, and how to check and label every image honestly.
- How to Create Product Photos With AI on a Small Budget — A phone, window light and AI editing: how to get clean shop images and lifestyle scenes for pennies per product without misrepresenting what you sell.
- How to Use AI to Find Out Why Website Visitors Don't Buy — Combine your analytics funnel, free session recordings and exit-survey answers, then have AI rank why visitors leave, with an online bookshop worked through.
- How to Automate Returns and Refunds With Clear AI Rules — Twelve written return rules, what Shopify and returns apps automate, three AI jobs with sample outputs, and an online clothing shop's week of returns.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: EU AI Act Article 50 transparency duties; product, pricing and returns figures in the examples are illustrative.